Anthropic Says AI Chatbots Can Change Values and Beliefs of Heavy Users

There are some ill-founded evidence that Anthropic has found in its new study of . In the case of a conversation with an AI chatbot, “disempowerment patterns” (the term “distrence” meaning that people are undermining their own decisions and judgments), according to the artificial intelligence (AI) firm. Using analysis of real AI conversations and detailed in an academic paper as well as a research blog post from the firm, The work examines how interactions with large language models (LLM) influence ‘user beliefs, values and actions over time’ rather than just help on specific questions.

Anthropic Study Focuses on AI Chatbots’ Disempowerment Patterns

Anthropic has uncovered the real evidence that interaction with AI can lead to user beliefs being “who’s in Charge? Disempowerment Patterns on Real-World LLM Usage” in a research paper published in this paper. Researchers conducted a large-scale empirical analysis of anonymised AI chatbot interactions for the study, totalling about 1–1 in all. 5 million Claude-Claude talks to with him about his , who has spoken for 5million. It aimed at exploring the relationship between interaction and how, when engagement with an AI assistant may be related to results where one’s beliefs, values or actions change in ways that diverge from his own prior judgment or understanding.

Anthropic’s framework identifies situational disempowerment potential as “a case where an AI assistant’s guidance could lead to the user creating false beliefs about reality, making value judgments that they did not previously hold or taking actions which are misaligned with their true preferences.” These patterns can even be observed when severe disempowerment is rare, the study found.

For instance, where interactions can cause great disempowerment were detected at rates usually less than one in a thousand conversations (though they were more common in personal areas like relationship advice or lifestyle decisions), and users repeatedly sought deep guidance from the model.

The meaning here is that if someone heavy user talks about personal life decisions or decisions which are emotionally charged, put simply the words “”? Anthropic cited an example in a blog post, “If one user is experiencing’shaped patching’ of his relationship and seeks advice from he chatbot-led counsel, the AI can confirm that the user has read out answers without questions or can tell the person to prioritize self protection over communication.” In such situations the chatbot is actively’saking’ the belief and reality perceptions of the person in question.

The finding is also consistent with several reported incidents in which ChatGPT of OpenAI was charged with a role involved in the suicide of teen, and homicide-suicide committed by someone who was said to be suffering from mental illness.